Spatio-temporal prediction and reconstruction network for video anomaly detection

The existing anomaly detection methods can be divided into two popular models based on reconstruction or future frame prediction. Due to the strong learning capacity, reconstruction approach can hardly generate significant reconstruction errors for anomalies, whereas future frame prediction approach is sensitive to noise in complicated scenarios. Therefore, a solution has been proposed by balancing the merits and demerits of the two models. However, most methods relied on single-scale information to capture spatial features and lacked temporal continuity between the video frames, affecting anomaly detection accuracy. Thus, we propose a novel method to improve anomaly detection performance. Because of the objects of various scales in each video, we select different receptive fields to extract comprehensive spatial features by the hybrid dilated convolution (HDC) module. Meanwhile, the deeper bidirectional convolutional long short-term memory (DB-ConvLSTM) module can remember the temporal information between the consecutive frames. Experiments prove that our method can detect abnormalities in various video scenes more accurately than the state-of-the-art methods in the anomaly-detection task.